On Fri, Jan 23, 2009 at 9:11 PM, John Van Pelt <[email protected]> wrote: > I *THINK* Eugene was asking for ideas that would be useful for > over-the-board estimation strategy, not a code-based approach. (Steven's > suggestion about O and G may help.) > > Another question might be: are Quackle's superleaves known to be all that > great -- or at any rate, how far from the mean do you have to be before > you're beyond the margin of error for incomplete word knowledge, incomplete > positional vision, and suboptimal strategy (assuming over-the-board > human-v-human scenarios)? In other words -- it's probably worth it to > "minimize" the error, but how close is it really worth the effort to get?
There is no reason to believe that optimal rack leave evaluations for games between opponents with perfect word knowledge would be the best evaluations to use in games between human opponents with less than perfect word knowledge. I recall a story where the rack leave value of C derived from Maven playing itself thousands of times proved to be too high when Maven played people (because people tend to play more closed which turned holding Cs from a bingoing asset into a tactical disadvantage). Steve > > -jvp
